AMD - Piledriver chitchat

[scaryjim]

You generally can't drop the minimum multiplier for idle states, but I've yet to find an AMD chip that won't let you lock the highest multiplier setting as low as 800MHz (be that x4 or x8, depending on platform), which is the default idle clock. I've currently got an A6-3500 locked to 1.5GHz max and undervolted to run as a web server (really must get round to finishing my mates website now ).

If you're running with C'n'Q then actually undervolting *does* make a difference to idle, because you can enable C'n'Q and it'll reduce the voltage at idle by the default offset: my Sempy used to run at 0.75v at idle, and it saved around 10W. And even with sections gated you still get leakage and you still have active parts to the CPU, which undervolting will (well, should) still reduce. I've yet to hit any obvious problems, anyway, with either my old sempy* or the new server.


* actually, that's not quite true - I did get a little instability on my Sempy, but only when I had it both unlocked to an Athlon X2 *and* undervolted. Which isn't that surprising

[watercooled]

The Sempron isn't really comparable to modern CPUs as it won't utilise gating etc. Modern CPUs idle down in the low-single-digits Watts when idle, when measured before the VRM. Leakage isn't really a concern if the core is gated. Some of the uncore will remain active but it doesn't use core voltages/clocks anyway.

My concern for instability is *because* of offset undervolting combined with frequency scaling (CnQ).

[CAT-THE-FIFTH]

Of course, if someone wants to donate a Richland A10 and a motherboard I'll be happy to do some investigation

If AMD is reading this thread,donate me one as well please!!

It would be interesting to see of an A10 6700T is doable in that small mini-ITX build I did a while back.

[teppic]

I've got a Richland A10 sitting in its sealed box, courtesy of AMD warranty. Waiting on the FM2+ boards

[scaryjim]

The Sempron isn't really comparable to modern CPUs as it won't utilise gating etc. Modern CPUs idle down in the low-single-digits Watts when idle, when measured before the VRM.

As I said, if I had a suitable CPU I'd happily test undervolting - idle draw has been coming down over the last couple of generations, so it's possible that undervolting wouldn't help at idle any more. That said, if core voltage was irrelevant to idle power draw I don't see why companies would still have core voltage reduce in idle situations - presumably in most situations some parts of the core remain ungated, and the reduced voltage would still lower the draw of those parts (although the percentage reduction may be less than I saw with the sempy).

Similarly, I'd imagine that if there is a stability problem for a CPU with minimal load at very low voltage that would show up fairly quickly under normal usage. Ideally you'd be able to adjust both the maximum load voltage and the minimum idle voltage (or the C'n'Q offset) so you could compensate for undervolting by reducing the voltage drop at idle. But I suspect that very few, if any, motherboards give you that option...

[watercooled]

Core voltage is reduced based on clock because there are more than two power states; something like basic web browsing might only wake the CPU core but keep it at the lowest (idle) clocks, or maybe a notch up. There is no clock when the core/clock is gated, so reducing it wouldn't make any difference. Mobile CPUs tend to be more aggressive here however, and have more granular control over scaling. The uncore runs on different voltage planes to the cores.

I recall seeing someone testing it, and the saw no difference at idle (I'll try to find it in a sec and post back here). It's partly why those more expensive, 'efficient' editions of chips of of questionable value unless you're thermally/power bound, as from tests I've seen they're often less efficient than the standard parts due to the increased processing time.

I played with some Phenom-specific clocking utility a while back which did actually allow you to modify voltages/clocks for each notch IIRC - the values are stored in BIOS, just usually not directly editable in the motherboard's UI.

Edit: That utility is called K10stat.

[DanceswithUnix]

I presume the point of these low power parts is for use with passive cooling cases and pico power supplies? If you just want quiet, then a big fan on a cheaper high end part would be better and as watercooled said that is more energy efficient too.

I notice the 100W parts have a 400MHz turbo, the 65W parts have a 600MHz turbo. I wonder if you get to keep that if you down clock to limit the power usage?

So many things that it would be nice to try, but I think my daughter would object if I started experimenting with her 5800K

Edit to add: Hmm, 6800K only turbos by 300MHz. Might want to try and cap the graphics part clocks as well.

[watercooled]

This looks like a fairly decent article on undervolting and proper testing for each state, with links to some suitable software: http://www.makeuseof.com/tag/how-und...-battery-life/

It will take time, but it's better than just nudging the offset down, stress testing at the max notch and hoping for the best IMO, potentially risking silent corruption etc.

[mikeo01]

Just to add some testing I've done with my Phenom, I actually use AMDMsrTweaker (similar to K10stat, but CLI based) and found reducing the voltage to a minimum of 0.7v at idle, whilst down clocking the cores have "some" effect on power usage.
The stock P-states set up would eat around 20-30w more than a custom state. However these only effect the register values.

I am also waiting for Kaveri but will probably pick up a Richland before they are released. K10stat works with Richland so I may be giving that ago before next year

[watercooled]

Register values?

[mikeo01]

In the system registry for the drivers.

[watercooled]

P-state parameters i.e. clock/voltage are defined in BIOS, although they're usually not editable in the BIOS menu, and the drivers/OS chooses what states to use at a given time. Or is that what you mean?

I'm probably just completely missing the point, but I don't get what you meant by that phrase?

[mikeo01]

The latter PhenomMSRTweaker, AMDMsrTweaker and I think K10stat can all use the Cool 'n' Quiet feature on the fly through the driver. Using offset voltage you can define the amount of voltage required by essentially "masking" the reported VID to a different set VID. For example I can set my multiplier as low as 0.5x and under the stock voltage by using a combination of offset and set VID to lower my power consumption.

That's why I think PhenomMSRTweaker had issues with CNQ enabled, because it tried to change the driver's settings on the fly. I don't know the technical details, that's all hash editing not editing the P-states through the BIOS, but at OS load.

[watercooled]

The p-state configuration still needs to be applied to the BIOS, but yeah it looks like that utility aims to replace the standard CnQ driver with its own, but based on how it works, it may not be suitable for Vista+ (better stock handling of CnQ), or for newer CPUs as it presumably doesn't account for recent power-saving features like individual core gating.

[CAT-THE-FIFTH]

This Jaguar based netbook looks interesting:

http://www.cnet.com/laptops/dell-ins...-35827688.html

It has quite a large battery for the price.

[mikeo01]

The p-state configuration still needs to be applied to the BIOS, but yeah it looks like that utility aims to replace the standard CnQ driver with its own, but based on how it works, it may not be ii suitable for Vista+ (better stock handling of CnQ), or for newer CPUs as it presumably doesn't account for recent power-saving features like individual core gating.

Well I'm using AMDMsrTweaker on Windows 8. Also it seems the utility applies the values to the drives settings. Since its CLI based it simply replaces the drives registry values. Once I've applied the values I do not need the utility. I have to actually uninstall the CPU from device manager to get the stock p-states back. So not entirely sure how it works.

Also it seems the utility is going to support Kaveri. On the forum the owner has apparently taken Llano and Richland into consideration.

http://www.xtremesystems.org/forums/...-AmdMsrTweaker